Bottom Line Studio AI
The agent bill nobody budgeted: retries
Agent budgets are written per run. The bill arrives per attempt. Here’s where the gap comes from, and the three guardrails that close it.
An agent never gets tired of trying again. Your budget does.
Sound familiar? The pilot cost almost nothing. Then usage went up, and the bill went up faster than usage. Nobody can say which runs were the expensive ones, or why.
The budget was per run. The bill was per attempt.
ICONIQ’s 2026 State of AI report has a line every finance lead should read: a workflow budgeted at ten cents a run drifted past a dollar fifty as the agents retried and corrected themselves. Nothing broke. The agent did exactly what it was built to do: when a step failed, it tried again.
Retries are where the money hides, because a retry usually isn’t a fresh, small call. The agent carries everything it has read so far into the next attempt: the instructions, the tool results, the answer that didn’t work. The fifth try costs more than the first, and the total for the run grows much faster than the number of tries.
Most teams aren’t watching the meter
Gartner surveyed 353 data and AI leaders at the end of last year. Only one in five worried that uncertain costs would limit AI value, and only 44% had financial guardrails or AI FinOps practices in place. That’s a lot of agents running on the honor system.
A meter beats a cheaper model
On a client engagement this year I built an account score that triggered paid enrichment and LLM-written briefs for an outbound team. The first version would have enriched everything it touched. Two rules changed that:
- Spend gated by the score. Enrichment and briefs only ran for accounts scoring 40 or above.
- A 30-day queue. Nothing got enriched twice inside the window.
Per-account enrichment cost went from about $0.50 to cents. Same model, same data vendor. The savings came from deciding where not to spend before anything ran.
Three guardrails, three different leaks
Each one stops a different kind of waste, and each has a cost you should choose on purpose.
| Guardrail | What it stops | What it costs you |
|---|---|---|
| Gate the spend | Paying for records that can’t move | You need a score you trust |
| Cap the retries | Runaway loops on a single run | Some runs stop and go to a person |
| Dedupe with a window | Paying twice for the same answer | Answers can go stale; pick the window |
Put the meter in before you scale
- Budget per run, report per attempt. Log attempts and tokens on every run so the expensive ones have names.
- Cap retries and hand off. After a set number of tries, the run stops and a person looks at it.
- Gate spend on a signal you trust. The expensive calls only run where they can change something.
- Dedupe on purpose. Choose the window by how fast the answer goes stale, not by habit.
- Review cost per outcome. Cost per call looks fine right up until the invoice.
Do you know what your most expensive agent run cost last week, and why?